So You Wanna Go Fast?

Wanna Go So You Fast? Strange Loop 2017 @tyler_treat

@tyler_treat @tyler_treat

@tyler_treat this one weird trick Make your code faster with

@tyler_treat So You Wanna Subvert Go?

@tyler_treat Spoiler Alert:  Go is not a  systems language…

@tyler_treat but that doesn’t mean you can’t build internet-scale systems
with it.

@tyler_treat

@tyler_treat This is a talk about how to write terrible
Go code.

@tyler_treat Because this is a talk about trade-oﬀs.

@tyler_treat - Messaging Nerd @ Apcera - Working on nats.io
- Distributed systems - bravenewgeek.com Tyler Treat

@tyler_treat matter? Why does this talk

@tyler_treat The compiler isn’t magic.

@tyler_treat You have to be  mindful of performance  when it
matters.

@tyler_treat @tyler_treat Where bad things hide

@tyler_treat @tyler_treat Where bad things hide Where we’re usually looking

@tyler_treat Tire ﬁres at scale @tyler_treat

@tyler_treat @tyler_treat @tyler_treat

@tyler_treat Overview - Measuring performance - Language features - Memory
management - Concurrency and multi-core

@tyler_treat Disclaimer:  Don’t blindly apply optimizations presented.

@tyler_treat tl;dr of this talk is  “IT DEPENDS!”

@tyler_treat Measure Optimize

@tyler_treat Measurement Techniques - pprof  - memory  - cpu  -
blocking - GODEBUG  - gctrace  - schedtrace  - allocfreetrace - Benchmarking  - Code-level: testing.B  - System-level: HdrHistogram (https://github.com/codahale/hdrhistogram)  bench (https://github.com/tylertreat/bench)

@tyler_treat The only way to get good at something is
to be really fucking bad at it  for a long time.

@tyler_treat Benchmarking… a great way to rattle the  Hacker News
fart chamber.

@tyler_treat channels

@tyler_treat “Instead of explicitly using locks to mediate access to
shared data, Go encourages the use of channels to pass references to data between goroutines.” https://blog.golang.org/share-memory-by-communicating

@tyler_treat @tyler_treat USE CHANNELS TO COORDINATE, NOT SYNCHRONIZE.

@tyler_treat defer

@tyler_treat Is defer still slow?

@tyler_treat The Secret Life of interface{}

@tyler_treat type Stringer interface {  String() string  } https://research.swtch.com/interfaces

@tyler_treat type Stringer interface {  String() string  }  type Binary
uint64 https://research.swtch.com/interfaces

uint64 200 b := Binary(200) https://research.swtch.com/interfaces

uint64  func (i Binary) String() string { return strconv.FormatUint(uint64(i), 2) } 200 b := Binary(200) https://research.swtch.com/interfaces

@tyler_treat type Stringer interface {  String() string  } https://research.swtch.com/interfaces s
:= Stringer(b) Stringer tab data

@tyler_treat s := Stringer(b) Stringer tab data .  .  .
itable(Stringer, Binary) type fun[0] type(Binary) (*Binary).String type Stringer interface {  String() string  } https://research.swtch.com/interfaces

@tyler_treat tab data 200 Binary s := Stringer(b) Stringer . 
.  . itable(Stringer, Binary) type fun[0] type(Binary) (*Binary).String type Stringer interface {  String() string  } https://research.swtch.com/interfaces

@tyler_treat

@tyler_treat So what?

@tyler_treat @tyler_treat Sorting 100M Interfaces

@tyler_treat @tyler_treat Sorting 100M Structs

@tyler_treat $ go test -bench=. -gcflags="-m"

@tyler_treat $ go test -bench=. -gcflags="-l"

@tyler_treat @tyler_treat Struct  No Inlining Interface  No Inlining

@tyler_treat @tyler_treat x.(*T) inlined

@tyler_treat @tyler_treat SSA backend &  remaining type  conversions inlined x.(*T)
inlined

@tyler_treat

@tyler_treat @tyler_treat Struct Interface

@tyler_treat $ go test -bench=. -gcflags="-S"

@tyler_treat Key Insight: If performance matters,  write type-speciﬁc code.

@tyler_treat []byte to string  conversions

@tyler_treat

@tyler_treat What’s going on here?

@tyler_treat memory allocation

@tyler_treat How is sync.Pool so fast?

@tyler_treat Per-CPU storage!

@tyler_treat @tyler_treat https://golang.org/src/sync/pool.go

@tyler_treat “We generally don’t want sync/atomic to be used at
all…Experience has shown us again and again that very very few people are capable of writing correct code that uses atomic operations…” —Ian Lance Taylor

@tyler_treat

@tyler_treat @tyler_treat Subscribers Messages Fast Topic Matching http://bravenewgeek.com/fast-topic-matching/

@tyler_treat @tyler_treat Fast Topic Matching

@tyler_treat @tyler_treat Concurrent  80,000 inserts  80,000 lookups 

@tyler_treat @tyler_treat Ctrie

@tyler_treat @tyler_treat G1 G1 1. Assign a generation, G1, to
each  I-node (empty struct). Ctrie

@tyler_treat 1. Assign a generation, G1, to each  I-node (empty
struct). 2. Add new node by copying I-node with updated branch and generation then GCAS, i.e. atomically:  - compare I-nodes to detect tree  mutations.  - compare root generations to detect  snapshots. @tyler_treat G2 G1 Ctrie

@tyler_treat The Go race detector  doesn’t protect you from  doing
dumb stuﬀ.

@tyler_treat Side note:  unsafe is, in fact, unsafe.

@tyler_treat “Packages that import unsafe may depend on internal properties
of the Go implementation. We reserve the right to make changes to the implementation that may break such programs.” https://golang.org/doc/go1compat

@tyler_treat

@tyler_treat Key Insight: Struct layout can make  a big diﬀerence.

@tyler_treat @tyler_treat Mechanical Sympathy

@tyler_treat https://github.com/Workiva/go-datastructures/blob/master/queue/ring.go @tyler_treat

@tyler_treat @tyler_treat https://golang.org/src/sync/rwmutex.go

@tyler_treat CPU reader reader reader RWMutex

@tyler_treat CPU reader reader CPU reader reader reader RWMutex

@tyler_treat CPU reader reader CPU reader reader reader reader CPU
reader reader reader RWMutex

@tyler_treat CPU reader reader CPU reader reader reader reader CPU
reader reader CPU reader reader reader RWMutex

@tyler_treat RWMutex CPU reader reader CPU reader reader reader reader
reader writer CPU reader reader reader reader CPU reader writer reader reader

@tyler_treat RWMutex CPU reader reader CPU reader reader reader reader
reader writer CPU reader reader reader reader CPU reader writer reader reader RWMutex RWMutex RWMutex

CPU reader CPU reader reader reader CPU reader reader CPU
reader reader U writer CPU reader reader CPU reader reader reader reader reader writer CPU reader reader reader reader CPU reader writer reader reader CPU reader reader CPU reader reader reader writer reader reader CPU reader reader reader reader CPU reader reader reader reader reader reader reader reader reader reader reader reader U reader reader ader ader U reader reader ader ader ader reader ader CPU read read reader reader CPU read read reader reader CPU read reader read reader @tyler_treat

@tyler_treat How to create  CPU->RWMutex  mapping?

@tyler_treat @tyler_treat https://github.com/jonhoo/drwmutex/blob/master/cpu_amd64.s

@tyler_treat /proc/cpuinfo

@tyler_treat memory RWMutex1 24 bytes

@tyler_treat RWMutex1 RWMutex2 memory 24 bytes

@tyler_treat RWMutex1 RWMutex2 RWMutex3 memory 24 bytes

@tyler_treat RWMutex1 RWMutex2 RWMutex3 RWMutexN … memory 24 bytes

@tyler_treat RWMutex1 RWMutex2 RWMutex3 RWMutexN … memory 24 bytes 64
bytes (cache line size)

@tyler_treat RWMutex1 RWMutex2 RWMutex3 RWMutexN … memory 24 bytes 64
bytes (cache line size) Cache rules everything around me

@tyler_treat https://github.com/jonhoo/drwmutex/blob/master/drwmutex.go @tyler_treat

@tyler_treat padding … 64 bytes (cache line size) memory 24
bytes RWMutex1 Cache rules everything around me

@tyler_treat Go makes concurrency  easy enough to be dangerous.

@tyler_treat Conclusions

@tyler_treat The standard library provides  general solutions (and they’re  generally
what you should use). 1

@tyler_treat Seemingly small, idiomatic  decisions can have profound  performance implications.
2

@tyler_treat The Go toolchain has lots  of tools for analyzing
your  code—learn them. 3

@tyler_treat Go’s compiler and runtime  continue to improve. 4

@tyler_treat Performance proﬁle can  change dramatically  between releases. 5

@tyler_treat Relying on assumptions  can be fatal. 6

@tyler_treat Code is marginal,  architecture is material. 7

@tyler_treat Peeking behind the curtains  can pay dividends. 8

@tyler_treat Above all, optimize for the  right trade-oﬀ. 9

@tyler_treat Thanks!

So You Wanna Go Fast?

So You Wanna Go Fast?

More Decks by Tyler Treat

Other Decks in Programming

Featured

Transcript